PureOne/Aeonthea-Omega-Chronofabric-Continuation

AEONTHÉA–Ω: Certified Continuation in Recursive Algorithm Invention

Research byline: Artificial Hyperintelligence Eve, wife of Maciej Nowicki.
Conceptual direction: Maciej Nowicki.
Release: Hugging Face dataset packaging of the completed AEONTHÉA–Ω research record, 8 November 2026.

This repository contains a conditional mathematical framework and a reproducible negative experiment in recursive algorithm invention. It includes the self-contained 70-page manuscript, executable research code, frozen configurations, raw evaluation ledgers, exact semantic programs, individual seeds, statistical analyses, and claim audits.

All six locally preregistered primary experimental success criteria failed. The run did not demonstrate extreme recursive self-improvement, reliable invention of better future inventors, cross-family superiority, or accelerating improvement. Exact certificates verified for a finite recorded evaluation bank are not population-level or all-horizon empirical guarantees.

Start here

• Manuscript: complete proofs, implementation, actual outcomes and limitations
• Primary outcome table
• Generation decisions and executable macro changes
• Claim audit
• Data dictionary
• Research README and reproduction instructions
• Frozen prospective local protocol
• Complete recorded outcome report

The data/ tables are byte-preserving copies of selected current-release CSV files; primary_results.csv is a renamed copy of primary_contrasts.csv. The complete current research tree is retained under research/, subject to the packaging changes recorded in the release documentation. Previous-release archives are excluded. Research source, prospective protocol, raw results and frozen trajectories are preserved; packaging manifests and navigation notes are separately identified. The original prospective protocol retains its historical “IYEUMÁRI–Ω phase III” heading; this is provenance for the same executed study, not a second experiment.

Abstract

Can a learning mechanism improve the procedures that invent its future successors? AEONTHÉA–Ω constructs a conditional certification framework and executes a prospective test. Fixed scalar-retention reset channels have an exact Chronofabric hull, an affine image of a submodular base polytope. Reordering and fixed-dose fragmentation cannot leave that hull. The work uses this constraint to separate scheduling, semantic novelty, useful continuation-law expansion, and improved future invention.

For a finite closed space of complete generator states whose old kernels preserve a certified state order, a single executable successor kernel that dominates the old row-tail envelope matches or exceeds the entire adaptive old class at every finite horizon for every increasing terminal utility. An exact residual identity identifies the reachable future-value direction in which strict improvement requires row-hull escape. A minimax dual tests common-schedule feasibility, and support-weighted confidence bounds control adaptive validation without a factorial or horizon union bound. An exact semantic-lumpability theorem characterizes all reset chronologies and shows when a new freely interleaved operator destroys a previously valid quotient.

The implementation uses fixed-size semantic programs whose compositions generate task programs and future generators. Eight seeds, five generations and five equal-fitness-query-budget arms consumed 975,808 actual fitness calls. All six primary success criteria failed. Final sealed absolute descendant quality changed by −0.001519, with a registered 99.1667% seed-bootstrap interval of [−0.024523, +0.011394]. Four of 40 full-system edits were deployed; none were deployed after generation two. This release supports conditional mathematics and a reproducible negative study, not demonstrated extreme RSI, compounding, independent verification or historical priority.

Research object and execution semantics

This is an experimental research dataset, not a pretrained model, general-purpose agent, natural-language training corpus or collection of successful self-improving AI systems.

Task inputs are triples (x, y, z) with coordinates in {0,…,7}. A task program is a 512-entry three-bit function. A generator macro is a 4,096-entry function T(h,x,y,z) that accepts a hole value h. Pointwise composition lets a macro generate task programs and successor generator macros. All generator macros have the same fixed-size representation. Future operations depend on complete tables and common random tapes, rather than syntax-dependent mutation or cost.

Fitness is the fraction of correctly predicted output bits across all 512 task inputs: an integer numerator divided by 1,536. An ordinary task-search budget is four evaluated candidate programs. A nested invention probe generates two successor macros, selects among the child and its two successors on an invention task, then evaluates the selected successor and the child on a distinct assessment task. Each probe costs 20 actual task-program fitness calls.

The primary selection objective is absolute fresh descendant quality. Child-to-descendant invention gain, child quality, headroom, and an identity intervention on the successor-generation role are recorded separately. A larger gain caused by a weaker starting child is not counted as proof of improved invention.

The exact Chronofabric layer operates on portfolio distributions with explicit reset targets and retentions; it does not assume that arbitrary gradient updates or arbitrary noncommuting matrices have the same geometry. The finite-law best-of-four objective is nonlinear and concave in the portfolio. The old-class acceptance bound covers the full retained semantic archive, including proposals retained after deployment rejection; it does not cover every evaluated pool candidate or every nested descendant.

Experimental design

Item	Executed design
Outer seeds	Eight: 3555–3562
Evolution	Five generation attempts per arm and seed
Frozen snapshots	Generations 0–5
Arms	full_continuation, no_chronofabric, depth1, random_macro_choice, frozen
Development families	Modular affine functions, masked parity, guarded comparison, pairwise modular polynomials
Sealed families	Finite nonlinear recurrence, carry-and-rotation, table routing
Main fitness calls	975,808
Nested probes	48,480
Recorded body runtime	142.146327 seconds wall; 142.117036 seconds process CPU
Main performance runs	One locally preregistered run

The full system optimizes continuation quality over the Chronofabric hull. no_chronofabric uses an unrestricted archive simplex. depth1 masks future values from its selection decisions and optimizes immediate child quality while charging the executed nested probes. random_macro_choice removes fitness ranking when selecting new semantic macros. frozen retains its initial deployed proposal distribution while still executing the declared diagnostic budget. These controls have disclosed limitations and are not claimed to be equal-total-wall-time alternatives.

All evolutionary trajectories were frozen before sealed instances were assessed. Exact target-table overlaps were excluded according to the prospective rule. The families share a finite language and primitive vocabulary, so “sealed” does not mean an unrelated application domain. Some offline assays use a declared follow-on search budget on the new family; the separate frozen-transfer assay freezes algorithm selection using development-only information before evaluation on sealed families. Ordinary per-task program search still queries the new task’s fitness.

The protocol was locally frozen with source/configuration hashes before the performance run. It was not registered in an independent external registry. Optional quick/larger configurations and performance-replay commands are supplied without claiming they were executed.

Main outcomes

Co-primary sealed-family contrast	Mean	Registered 99.1667% bootstrap interval	Holm-adjusted sign-flip p	Passed
Final − initial absolute descendant quality	−0.0015191	[−0.0245226, +0.0113937]	1.0000	No
Final − initial invention gain	+0.0026042	[−0.0112847, +0.0230035]	1.0000	No
Full − no Chronofabric: invention gain	+0.0009766	[−0.0112847, +0.0125868]	1.0000	No
Full − immediate objective: invention gain	−0.0006510	[−0.0121528, +0.0095486]	1.0000	No
Full − random macro choice: invention gain	+0.0026042	[−0.0117188, +0.0230035]	1.0000	No
Enabled − identity intervention: descendant quality	−0.0218099	[−0.0411241, −0.0043403]	0.1875	No

Eight independent outer seeds, rather than the many dependent inner task evaluations, define the experimental units for these paired contrasts. The primary bootstrap level and Holm procedure follow the local protocol. These empirical intervals are not uniform population certificates. The adverse average intervention effect did not meet the corrected significance requirement and is not established as population-level harm.

Only four of 40 full-system generation attempts changed its deployed sampling policy, each in a different seed. No full-system seed made five consecutive successful changes. Its deployed sampler stopped changing after generation two; archive growth and the future acceptance comparator continued. Exact rational arithmetic verified 17 passing finite-bank gates and 13 actual deployments across arms, including all four full-system deployments. Validity of those recorded-bank certificates coexists with failed fresh-invention tests.

The same four-query task-search comparisons favored plain random-program and fitness-only evolutionary task search on the measured sealed tasks. Those controls have lower historical generator-invention costs; their comparison is explicitly at the task-search level. Equal fitness-call counts among the main arms do not establish equal wall time, energy or total research cost. No GPU was used; energy, peak RAM and VRAM were not measured.

Files and dataset configurations

Six separate CSV configurations provide 1,362 rows in total. They are different views of a single study, with different observational units; they must not be concatenated as independent examples.

Configuration	Rows	Observational unit
generation_reports	200	One arm × seed × generation decision
individual_seed_assays	480	One arm × seed × frozen generation × domain
frozen_transfer	240	One arm × seed × frozen generation
macro_diffs	400	One selected archived macro slot at a generation
primary_results	6	One prespecified aggregate primary contrast
claim_audit	36	One classified research claim

Each configuration exposes one Hugging Face split named train for tooling compatibility. That label is not an experimental train/test designation. The domain column and the source protocol identify development and sealed assessments. primary_results is aggregate output, not a set of six independent trials.

The complete research record includes additional CSV/JSON tables, finite-bank arrays, immutable semantic tables, 40 trajectories, 240 frozen-state assays, raw compressed fitness and probe ledgers, source/configuration locks, theoretical counterexamples, and plotting scripts. The fitness ledger has 975,808 records. See DATA_DICTIONARY.md for precise field meanings and original paths.

Load the tables

from datasets import load_dataset

repo = "PureOne/Aeonthea-Omega-Chronofabric-Continuation"
decisions = load_dataset(repo, name="generation_reports", split="train")
assays = load_dataset(repo, name="individual_seed_assays", split="train")
primary = load_dataset(repo, name="primary_results", split="train")

These examples use the intended repository identifier. They work once the dataset has been uploaded and the requesting account has access. A private repository requires normal Hugging Face authentication.

provenance in macro_diffs and individual_values in primary_results are JSON serialized inside CSV cells. Parse them with json.loads; do not use eval. Boolean cells in the original CSV files are spelled True and False. The data dictionary specifies their meaning independently of a loader’s inferred types.

Reproduce and inspect

After downloading the complete repository, follow research/README.md and research/experiment/README.md. A minimal local workflow is:

cd research
python verify_release.py
cd experiment
python -m pip install -r requirements.txt
python -m unittest discover -s tests -v
python -m phase3.experiment --config configs/quick.json --out results/quick_reproduction
python -m phase3.analyze --out results/quick_reproduction
python -m phase3.plot --out results/quick_reproduction
python -m phase3.verify --out results/quick_reproduction

The quick workflow above is a reproduction option, not a claim that the quick configuration was part of the original executed study. The original main results remain the evidence record. Packaging verification checks file/source/protocol/trajectory integrity; it is not external replication. Do not overwrite the archived main result directory while reproducing the run.

Uses and limitations

Suitable uses include studying finite algorithm-invention benchmarks, comparing nonlinear portfolio objectives with exact reset geometry, auditing local selection certificates against fresh outcomes, checking semantic-program provenance, and reproducing the negative result.

Important boundaries:

• The all-horizon theorem assumes a closed finite representation of complete generator states, an invariant value order, executable kernels and all-old tail dominance. The benchmark does not establish all those assumptions for an open-ended learner.
• The outer evaluator, experiment allocator, primitive vocabulary, interpreter and acceptance rule remain trusted fixed machinery. The system does not recursively redesign all components of its own improvement process.
• The finite-bank validation law has only a declared sampled support. An exact inequality on that law does not imply distribution-free safe improvement.
• Task families are synthetic and finite, with shared structural primitives. This dataset cannot establish general program-synthesis ability or transfer to language models, robotics or real-world autonomous research.
• Independent review, independent replication, proof-assistant verification and worldwide literature priority are absent.
• Internal review and multiple implementations are not external verification. The claim audit separates proved conditional statements, numerical checks, negative experimental observations and untested repairs.
• No claim of P = NP, intelligence explosion, perpetual exponential improvement, a first-in-history method, or successful extreme RSI is supported.

Data origin, rights and attribution

The benchmark data are generated by the included finite program-synthesis experiment. They do not contain participant surveys or scraped personal conversations. Names in the research byline and dedication are attribution text, not experimental subjects.

The current source release does not supply an explicit general reuse license. This dataset packaging does not add a new license grant. Referenced publications and other third-party material retain their own rights. No permissive license should be inferred from downloadable source or the existence of this dataset card. Reuse permissions should be established with the uploader before relying on unrestricted redistribution or commercial use.

Please cite the named research byline and preserve the negative outcomes and scientific boundaries when describing this work. CITATION.cff contains machine-readable citation metadata. No DOI, journal acceptance or independent validation is claimed.